ci: Customize benchmark PR comment action

[samuelburnham]

PR benchmark comment customization

This PR enables customization of the issue_comment PR comment action from directly within the !gpu-benchmark comment. We can now run multiple benchmarks, with customizable features and env vars. The initial format is as follows:

!gpu-benchmark
Benches:
fibonacci,end2end
Features:
cuda,portable
Env vars:
LURK_RC=100,600,900
LURK_BENCH_NOISE_THRESHOLD=0.05

Implementation notes

• Splits the benchmark into a setup action to parse the input parameters, then runs one GPU benchmark job per benchmark in a matrix job. This is due to the inherent limitation of boa-dev/criterion-compare-action only being able to run one benchmark or all of them (the latter with some hacks).
• Removes reliance on the bench.env file used by just in favor of user specification in the comment
• Removes the CPU-only benchmark as it's seldom used. Probably a better solution would be to turn off the cuda feature once implemented, as it's hardcoded here.

Limitations

• Brittle parsing format due to rudimentary awk script, which should be improved in future work.
• Benches and features must be provided on one line, separated by comma
• Env vars must be listed one per line

Successful run

https://github.com/lurk-lab/ci-lab/actions/runs/7580695577

[huitseeker]

I'm happy with exploring a richer set of benchmarks, but I really want to support a sensible default, as we do now.

There's a lot of friction brought by needing to learn what is the default benchmark, and exposure to janky parsing does not make the situation better (OTOH, for exotic needs, janky parsing is just fine).

Case in point: I'd like the PR to be backward compatible with the following "dumb" behavior:
argumentcomputer/ci-lab#59 (comment)
run:
https://github.com/lurk-lab/ci-lab/actions/runs/7670967389
error point:
https://github.com/lurk-lab/ci-lab/blob/main/.github/workflows/pr-comment.yml#L50

[samuelburnham]

Superseded by #1114